Google BigQuery SQL: Order two columns independently

Question

Say I have some data like:

grp   v1   v2
---   --   --
 2    5    7
 2    4    9
 3    10   2
 3    11   1

I'd like to create new columns which are independent of the ordering of the table - such that the two columns have independent orderings, ie sort by v1 independently of v2, while partitioning by grp.

The result (independently ordered, partitioned by grp) would be:

grp   v1   v2  v1_ordered v2_ordered
---   --   --  ---------- ----------
 2    5    7       4          7
 2    4    9       5          9
 3    10   2      10          1
 3    11   1      11          2

One way to do this is to create two tables and CROSS JOIN. However, I'm working with too many rows of data for this to be computationally tractable - is there a way to do this within a single query without a JOIN?

Basically, I'd like to write SQL like:

SELECT
  *,
  v1 OVER (PARTITION BY grp ORDER BY v1 ASC) as v1_ordered,
  v2 OVER (PARTITION BY grp ORDER BY v2 ASC) as v2_ordered
FROM [example_table]

This breaks table row meaning, but it's a necessary feature for many applications - for example computing ordered correlation between two fields CORR(v1_ordered, v2_ordered).

Is this possible?

Answer 1

I think you are in right direction! You just need to use proper window function . Row_number() in this case. And it should work!

Adding working example as per @cgn request:
I dont think there is way to totally avoid use of JOIN.
At the same time below example uses just ONE JOIN vs TWO JOIN s in other answers:

SELECT 
  a.grp AS grp, 
  a.v1 AS v1, 
  a.v2 AS v2, 
  a.v1 AS v1_ordered, 
  b.v2 AS v2_ordered 
FROM (
  SELECT grp, v1, v2, ROW_NUMBER() OVER(PARTITION BY grp ORDER BY v1) AS v1_order
  FROM [example_table]
) AS a
JOIN (
  SELECT grp, v1, v2, ROW_NUMBER() OVER(PARTITION BY grp ORDER BY v2) AS v2_order
  FROM [example_table]
) AS b
ON a.grp = b.grp AND a.v1_order = b.v2_order 

Result is as expected:

grp v1  v2  v1_ordered  v2_ordered   
2    4   9           4           7   
2    5   7           5           9   
3   10   2          10           1   
3   11   1          11           2   

And now you can use CORR() as below

SELECT grp, CORR(v1_ordered, v2_ordered) AS [corr]
FROM (
  SELECT 
    a.grp AS grp, 
    a.v1 AS v1, 
    a.v2 AS v2, 
    a.v1 AS v1_ordered, 
    b.v2 AS v2_ordered 
  FROM (
    SELECT grp, v1, v2, ROW_NUMBER() OVER(PARTITION BY grp ORDER BY v1) AS v1_order
    FROM [example_table]
  ) AS a
  JOIN (
    SELECT grp, v1, v2, ROW_NUMBER() OVER(PARTITION BY grp ORDER BY v2) AS v2_order
    FROM [example_table]
  ) AS b
  ON a.grp = b.grp AND a.v1_order = b.v2_order
)
GROUP BY grp

Answer 2

This will work for you.

SQLFiddle Demo in SQL Server

Note: The sequence you mentioned in the sample, is not necessary how the rows are returned from database. In my case, for v1 , I got 4,5,10,11 unlike your 5,4,10,11 . However, your output will be same as you wanted.

Select t.grp,t.v1,t.v2,
v1.v1 as v1_ordered,v2.v2 as v2_ordered
From
(
    select t1.*,
    row_number() over (partition by grp
                   Order by v1) v1o
    ,
    row_number() over (partition by grp
                   Order by v2) v2o
    from table1 t1
) t
Inner join
(
    Select t.*,
    row_number() over (partition by grp
                   Order by v1) v1o
    From table1 t
) v1
On t.grp=v1.grp
And t.v1o=v1.v1o
Inner join
(
    Select t.*,
    row_number() over (partition by grp
                   Order by v2) v2o
    From table1 t
) v2
On t.grp=v2.grp
And t.v1o=v2.v2o

Output:

+------+-----+-----+-------------+------------+
| grp  | v1  | v2  | v1_ordered  | v2_ordered |
+------+-----+-----+-------------+------------+
|   2  |  4  |  9  |          4  |          7 |
|   2  |  5  |  7  |          5  |          9 |
|   3  | 10  |  2  |         10  |          1 |
|   3  | 11  |  1  |         11  |          2 |
+------+-----+-----+-------------+------------+

Answer 3

AI'm not 100% sure this works in BigQuery, but here is goes:

select e.*, ev1.v1, ev2.v2
from (select e.*,
             row_number() over (partition by grp order by v1) as seqnum_v1,
             row_number() over (partition by grp order by v2) as seqnum_v2
      from example e
     ) e join
     (select e.*, row_number() over (partition by grp order by v1) as seqnum_v1
      from example e
     ) ev1
     on ev1.grp = e.grp and ev1.seqnum_v1 = e.seqnum_v1 join
     (select e.*, row_number() over (partition by grp order by v2) as seqnum_v2
      from example e
     ) ev2
     on ev2.grp = e.grp and ev2.seqnum_v2 = e.seqnum_v2;

The idea is to assign an independent ordering to each of the columns. Then join back to the original table to get the actual value.